Bank Exposures and Sovereign Stress Transmission*

Transcription

1 Review of Finance, 2017, doi: /rof/rfx038 Advance Access Publication Date: 17 August 2017 Bank Exposures and Sovereign Stress Transmission* Carlo Altavilla 1, Marco Pagano 2, and Saverio Simonelli 2 1 European Central Bank and CSEF and 2 Università di Napoli Federico II and CSEF Abstract Using novel monthly data for 226 euro-area banks from 2007 to 2015, we investigate the determinants of banks sovereign exposures and their effects on lending during and after the crisis. Public, bailed-out and poorly capitalized banks responded to sovereign stress by purchasing domestic public debt more than other banks, consistent with both the moral suasion and the carry trade hypothesis. Public banks purchases grew especially in coincidence with the largest ECB liquidity injections, which therefore reinforced the moral suasion mechanism. Bank exposures significantly amplified the impact of sovereign stress on bank lending to domestic firms, as well as on lending by foreign subsidiaries of stressed-country banks to firms in non-stressed countries. Altogether, our evidence connects this amplification effect and its cross-border transmission to the moral suasion exerted by domestic governments on banks during the crisis. * We thank the editor (Alex Edmans) and an anonymous referee for very helpful suggestions. Viral Acharya, Efraim Benmelech, Markus Brunnermeier, Elena Carletti, Charles Calomiris, Hans Degryse, Itamar Drechsler, Tim Eisert, Andrew Ellul, Nicola Gennaioli, Rony Hamaui, Balint Horvath, Luc Laeven, José Maria Liberti, Andrea Polo, Steven Ongena, Thomas Philippon, Philipp Schnabl, Harald Uhlig, and Luigi Zingales provided insightful comments. Useful input was also provided by participants in seminars at the Bank of Finland, Central Bank of Ireland, ECB, ETH Zurich, Goethe University (Frankfurt), Graduate Institute (Geneva), Humboldt University, LUISS, Macquaire University, National Bank of Belgium, UPF, UNSW, USI (Lugano), and at the following conferences: CSEF-CIM-UCL Conference on Macroeconomics after the Great Recession, CSEF-IGIER Symposium on Economics and Institutions, 2017 DNB-EBC-CPER Conference on Avoiding and Resolving Banking Crises, 2016 EEA and EFA meetings, ETH-NYU Conference on Governance and Risk-Taking, 2016 NBER Summer Institute, RELTIF workshop, eighth Summer Macro-Finance Workshop, 2015 RIDGE Workshop on Financial Stability, 2016 Riksbank Macroprudential Conference, SCE 23rd Conference on Computing in Economics and Finance, 2015 SIE conference, and Workshop on Systemic Risk, Financial Networks and the Real Economy (Milan). Part of the project was done while Saverio Simonelli was visiting the ECB. M.P. and S.S. acknowledge financial support from the CEPR/Assonime RELTIF Programme and EIEF. The opinions in this paper do not necessarily reflect the views of the European Central Bank and the Eurosystem. VC The Authors Published by Oxford University Press on behalf of the European Finance Association. All rights reserved. For Permissions, please journals.permissions@oup.com

2 2104 C. Altavilla et al. JEL classification: E44, F3, G01, G21, H63 Keywords: Sovereign exposures, Sovereign stress, Moral suasion, Carry trade, Bank lending, Crisis Received June 8, 2017; accepted July 3, 2017 by Editor Alex Edmans. 1. Introduction The euro-area debt crisis and its aftermath are a natural testing ground to assess the role of banks exposures in the transmission of sovereign stress to the credit market. In this paper, the evidence generated by the crisis is used to address two closely related research questions: first, how did banks change their public debt holdings in response to sovereign stress, and how did their response vary depending on their characteristics? Second, did their different sovereign exposures amplify the transmission of stress to their lending? To answer these questions, we draw on a unique data set covering 226 euro-area banks at monthly frequency from 2007 to Exploiting the heterogeneity in banks characteristics allows us to test competing hypotheses regarding the response of their sovereign exposures to sovereign stress. Furthermore, exploiting the bank-specific dynamics of exposures enables us to quantify their contribution to the transmission of sovereign stress to lending. We establish two main results. First, publicly owned and recently bailed-out banks reacted to sovereign stress by purchasing significantly more domestic public debt than other banks, and boosted their purchases especially at the time of the two large liquidity injections by the ECB in December 2011 and March Since public and recently bailed-out banks are more likely to yield to political pressure than other banks, the evidence is consistent with their public debt purchases during the crisis being driven by the respective government s pressure the moral suasion hypothesis. 1 The low funding costs due to the ECB liquidity injections appear to have reinforced this mechanism: the estimates imply that, at the time of these injections, stressed-country public banks increased their sovereign debt holdings by 17% more than private banks. We also find that stressed-country banks with low regulatory capital bought more domestic public debt than other banks, in line with the view that they engaged in yield-seeking behavior to gamble for resurrection the carry trade hypothesis. The two hypotheses appear to have about the same explanatory power and to apply to almost completely disjoint sets of banks in our sample. Second, stressed-country banks with larger sovereign exposures cut lending more deeply than less exposed banks when sovereign stress increased, and expanded lending more when sovereign stress abated. The granular nature of our data enables us to estimate precisely the amplification effect associated with sovereign exposures: a 1-standard-deviation drop in the price of government bonds reduced the loan growth of the median domestic bank by 1 This hypothesis is formalized by Uhlig (2013), who shows that fiscally vulnerable governments have an incentive to allow domestic banks to hold home risky bonds, in order to borrow more cheaply, while non-vulnerable governments will impose tighter regulation. Battistini, Pagano, and Simonelli (2014) argue that sovereign stress heightens this incentive, generating a positive relationship between sovereign yields and banks holdings of domestic debt, and refer to this prediction as the moral suasion hypothesis, a label also used in subsequent work.

3 Bank and Sovereign Stress Transmission percentage points, which is 20% of the standard deviation of loan growth. This amplification mechanism can account for the entire drop in lending by the average bank in stressed countries at the peak of the sovereign crisis, that is, between mid-2010 and mid In principle, domestic customers may reduce their demand for lending at times of sovereign stress, thereby introducing an omitted-variable bias in our lending regressions. The unconsolidated nature of our banks balance-sheet data helps us to address this endogeneity concern: we investigate whether losses on sovereign debt incurred by parent banks in stressed countries affected their foreign subsidiaries loans to firms in non-stressed countries, whose demand for credit should not respond to sovereign stress. The sovereign exposures of the parent banks turn out to affect the lending of their foreign subsidiaries, to an extent that is comparable to that found for lending to domestic firms by the respective parent banks. This indicates that our estimates of the amplification effect are not driven by demand-side factors. Beside addressing endogeneity concerns, these estimates have substantive economic implications: they show that banks sovereign exposures amplify the impact of sovereign debt repricing not only on their domestic but also on their foreign lending, and thereby contribute to the international transmission of sovereign stress. Another possible concern is that banks losses on sovereign holdings may not be exogenous in our lending regressions, for instance because banks with larger sovereign holdings have clients whose solvency is more sensitive to sovereign risk. To this purpose, we build on the previous findings that public ownership and bailout events are key determinants of banks sovereign exposures, and interact these variables with sovereign repricing to construct instruments for banks losses on sovereign holdings. The exclusion restriction required for the validity of these instruments is that the loans of public and bailed-out banks react differently to sovereign stress only because they have larger sovereign exposures: this restriction would be violated if the customers of public and bailed-out banks became riskier at times of sovereign stress. We show instead that for these banks the fraction of impaired loans does not increase more than for other banks at the time of sovereign stress, thus supporting the exclusion restriction. The instrumental variable (IV) regressions confirm the amplification effect of sovereign exposures on stressed-country bank lending. These IV estimates indicate that this amplification mechanism can be traced back to the moral suasion exerted by governments on banks during the crisis, underscoring the tight connection between the two research issues addressed by our analysis. Our paper is related to a large literature on the drivers of domestic sovereign exposures during sovereign crises. Indirect evidence on such drivers was first provided by Acharya and Steffen (2015), who document that the loadings of bank stock returns on sovereign debt returns are higher for low-capitalized and recently bailed-out banks. They interpret these findings as evidence for the carry trade and moral suasion hypotheses, respectively. This interpretation is warranted if factor loadings proxy for banks sovereign exposures, but not if these loadings were to reflect just banks dependence on public bailout guarantees: the stocks of less capitalized banks and recently bailout banks may be more sensitive to public debt returns simply because they depend more on the government as backstop. Instead, our month-by-month observations of banks sovereign holdings enable us to directly estimate the impact of sovereign stress on the portfolios of banks with different characteristics. Ongena, Popov, and van Horen (2016) find that stressed-country domestic banks bought more sovereign debt than foreign banks when the domestic government s financing needs were particularly high. De Marco and Macchiavelli (2014) report that banks with

4 2106 C. Altavilla et al. sizeable government ownership or politically appointed directors feature more home-biased sovereign portfolios than privately owned and managed banks. These findings are consistent with the moral suasion hypothesis. Instead, Buch, Koetter, and Ohls (2016) report evidence supporting the carry trade hypothesis using granular information on German banks. Finally, Horvath, Huizinga, and Ioannidou (2015) test both hypotheses, but in separate regressions, so that from their estimates it is unclear whether both would have explanatory power in a nested specification. Other papers investigate whether central bank liquidity fueled the purchase of sovereign debt by banks. Drechsler et al. (2016) document that less capitalized banks bought more domestic sovereign debt after the extraordinary liquidity provision by the ECB in December 2011 and March However, Peydro, Polo, and Sette (2017) find that more not less capitalized Italian banks bought high-yield bonds when monetary policy softened, countering the idea that liquidity injections encouraged banks carry trades. Ongena, Popov, and van Horen (2016) find that domestic and public banks engaged in larger sovereign debt purchases but these were not fueled by the ECB liquidity injections. In contrast with their evidence, we document that the ECB liquidity injections in 2011 and 2012 amplified the moral suasion channel, since they appear to have enabled public banks to buy more sovereign debt. Instead, we find no evidence that these liquidity injections reinforced the carry trade channel, by making poorly capitalized banks more inclined to buy stressed public debt. Our paper is also related to the literature on the transmission of sovereign stress to lending activity. Gennaioli, Martin, and Rossi (2014a) present a model in which sovereign defaults reduce private lending by undermining the balance sheets of domestic banks, the more so the larger their holdings of government debt, and test these predictions on crosscountry evidence; they also test them on bank-level data in a companion paper (Gennaioli, Martin, and Rossi, 2014b). Becker and Ivashina (2014) use company data on bank borrowing and bond issuance to show that European companies were more likely to replace bank loans with bond issues when banks in their country held more domestic sovereign debt and when that debt was risky. De Marco (2017) and Popov and Van Horen (2014) show that the euro-area banks with larger sovereign exposures in the EBA stress tests participated to the syndicated loan market less than banks with lower exposures, and raised their lending rates more sharply. 2 All these studies suffer from the lack of accurate time series of banklevel data for banks sovereign exposures. Gennaioli, Martin, and Rossi (2014b) rely on banks total bond holdings, which lump domestic government bonds together with nondomestic bonds. The other three studies use data on sovereign exposures drawn from the EBA stress tests, and thus refer only to (at most) four dates and to a small sample of systemically important banks. To identify the transmission of sovereign stress to lending via banks sovereign exposures, it is important to control for the demand for loans by firms. The recent contributions by Acharya et al. (2015) and Carpinelli and Crosignani (2017) achieve such identification following the methodology proposed by Khwaja and Mian (2008): they analyze the change in loans issued to the same firm by banks with different exposures to sovereign risk. In our study, we control for loan demand in other ways, since we do not have bank firm matched loan data. However, our data are more complete in terms 2 De Marco (2017) documents this finding also using yearly balance-sheet data on bank loans, besides syndicated loan data.

5 Bank and Sovereign Stress Transmission 2107 of coverage of banks, countries, and time, as they refer to a sample of banks providing about 70% of total euro-area lending, and track bank-level sovereign exposures and lending policies throughout the crisis and after its abatement, rather than at specific dates and for a segment of the credit market. In contrast, Acharya et al. (2015) measure bank lending with data on syndicated loans, which account for just 10% of total euro-area lending and cater mostly to large, established corporations, while Carpinelli and Crosignani (2017) focus only on Italian banks. The structure of the paper is as follows. Section 2 describes the data, illustrating the variation in bank-level exposures and presenting some stylized facts. Section 3 analyzes the determinants of banks domestic sovereign exposures. Section 4 examines whether these exposures influenced the impact of sovereign stress on bank lending. Section 5 concludes. 2. Data and Stylized Facts This section describes our data and sets out some stylized facts about euro-area banks holdings of domestic sovereign bonds and their relationship with bank lending. These not only help to gauge the correlations in the data at aggregate level but also point to the additional insights that can be gleaned from bank-level data. Our analysis is based on a unique, proprietary data set of balance-sheet items (BSI) at bank level (Individual Balance-Sheet Items or IBSI), which is regularly updated by the ECB. We use monthly observations on the main balance-sheet indicators (assets and liabilities) from June 2007 to February The sample contains a total of 226 unconsolidated banks in eighteen euro-area countries (Table I), the highest coverage being in the largest countries: Germany (sixty), France (thirty-two), Italy (twenty-four), and Spain (twentythree). The banks are observed at unconsolidated level: 119 group head banks, 49 domestic subsidiaries, and 59 foreign subsidiaries (some affiliated to UK or Danish groups). 3 These data are merged with data on bank share ownership from Bankscope and handcollected data about bailout dates from the EU Commission state aid database. The data include monthly observations of the benchmark 10-year and 5-year sovereign yields and survey-based consensus yield forecasts at 3-month and 12-month horizons. Yields for euroarea countries are drawn from Datastream; survey-based forecasts are from Consensus Economics and are available only for France, Germany, Italy, the Netherlands, and Spain. For details on data definitions and sources, see the Appendix. The representativeness of the sample is shown in Table II, which reports main assets (defined as total assets less derivatives), loans to non-financial corporations and holdings of government bonds for the banks in our data set as a fraction of the national aggregate, drawn from the ECB BSI database. On average, for the main variables our data cover about 3 Our analysis is based on the IBSI data release of April 15, 2015, which contained data for 252 banks. Of these, we removed twenty-six banks featuring one or more of the following: (i) less than 12 months of observations were available for loans and exposures; (ii) loans equal to zero for the entire sample (with at most sparse spikes); (iii) frequent and extreme jumps in exposures or loans. Of the removed banks, two are Finnish, five French, five German, two Irish, two Italian, five Latvian, one is from Luxembourg, one Slovenian, and three are Spanish. We also remove all negative values of domestic sovereign holdings, equity, main assets, and lending.

6 2108 C. Altavilla et al. Table I. Distribution of the banks by country and ownership For each country, the table reports the total number of individual banks and their breakdown according to the country in which they operate and domestic or foreign ownership. Total Domestic banks Foreign banks Head banks subsidiaries Austria Belgium Cyprus Estonia Finland France Germany Greece Ireland Italy Luxembourg Malta Netherlands Portugal Slovakia Slovenia Spain Total % of the corresponding country aggregate. The bottom row of the table shows that weighting country coverage by GDP does not change the results. Our data are far more representative of the euro-area banking system than those used in previous studies, along several dimensions. First, our sample has data for the sovereign exposures of 226 banks, compared with at most 91 banks in the pre-2014 EBA stress test data, and for 93 months, compared with the few snapshots of the EBA stress tests. Second, as illustrated in Table II, our bank loan data cover almost 70% of the corresponding national lending aggregates, compared with the 10% coverage of the syndicated loan data used by Popov and Van Horen (2014), De Marco (2017), and Acharya et al. (2015). Descriptive statistics for the main variables are shown in Panel A of Table III, and for bank characteristics in Panel B. As in the subsequent analysis, the statistics are computed separately for two groups of countries: stressed (Cyprus, Greece, Ireland, Italy, Portugal, Slovenia, and Spain) and non-stressed (Austria, Belgium, Estonia, Finland, France, Germany, Luxembourg, Malta, the Netherlands, and Slovakia). We define as stressed that is, subject to high sovereign stress countries whose 10-year sovereign yield exceeded 6% (or, equivalently, four points above the German yield) for at least one quarter in our sample period. Table III reveals that banks in these two groups of countries behaved quite differently in several respects. First, their domestic sovereign exposures (the ratio of government debt holdings to main assets) are greater in stressed countries (4.9%) than in non-stressed ones (3.8%), while the opposite applies to non-domestic euro-area exposures (1% versus

7 Bank and Sovereign Stress Transmission 2109 Table II. Sample representativeness For each country, the table shows the aggregate values of main assets, loans to non-financial corporations (NFCs) and holdings of government debt in our dataset in January 2015 as percentages of the same variables in the aggregate data reported in the BSI statistics of the ECB. Ratio of IBSI aggregates to BSI totals (%) Main assets Loans to non-financial corporations Bank holdings of sovereign debt Austria Belgium Cyprus Estonia Finland France Germany Greece Ireland Italy Luxembourg Malta Netherlands Portugal Slovakia Slovenia Spain Average Weighted average %). 4 Hence, in stressed countries the sovereign debt portfolios of banks are more home-biased than in non-stressed countries. (Unfortunately, we cannot measure the diversification of sovereign debt portfolios more precisely, because our data do not break non-domestic exposures down by sovereign issuer.) Second, banks accumulated domestic sovereign debt twice as fast in stressed as in non-stressed countries (2% versus 1% on a quarterly basis). Third, in stressed countries loans to firms are a larger fraction of bank assets than in non-stressed countries but grow less. However, in both groups of countries there is considerable dispersion in the sovereign exposures of banks, as well as in the growth of bank sovereign holdings and lending to firms. Sovereign exposures feature substantial variation both over time and crosssectionally: in the stressed countries, their overall standard deviation is 4.9%, the same 4 Banks sovereign holdings are partly at market prices and partly at book values. They are marked to market if the bank classes them in its trading book (i.e., either available for sale or held for trading ). They are at book values if the bank classes them in its banking book (i.e., held to maturity ). Our data do not contain the breakdown between these two components. In the forty-five euro-area banks present in the EBA stress test data, trading-book sovereigns account for 59% of the total for banks in stressed and 48% in non-stressed countries.

8 2110 C. Altavilla et al. Table III. Descriptive statistics The table presents the mean, median, and standard deviation of banks monthly sovereign exposures, loans to firms (Panel A), and characteristics (Panel B). The stressed countries are Cyprus, Greece, Ireland, Italy, Portugal, Slovenia, and Spain; the non-stressed countries are Austria, Belgium, Estonia, Finland, France, Germany, Luxembourg, Malta, the Netherlands, and Slovakia. Domestic sovereign exposures are domestic sovereign debt as a fraction of the corresponding bank s main assets. Bank lending is the bank loans to non-financial corporations as a fraction of the corresponding banks main assets. Bank lending growth and sovereign holdings growth are the quarterly growth rates (in percent) of bank loans to non-financial companies and of their sovereign holdings. Leverage ratio is the ratio of banks total assets to their equity capital. T1/RWA is the ratio of Tier-1 common equity to risk-weighted assets. Public is the fraction of banks shares owned by local or national government or publicly controlled institutions (Fondazioni in Italy, Fundaciones and Cajas in Spain, and Sparkasse and Landesbank in Germany). Bailout equals 1 starting in the quarter in which a bank was bailed out (unless acquired in the two subsequent quarters), and 0 before that date. Panel A. Domestic exposures, bank lending, and interest rates (%) Stressed countries Mean Median Standard deviation Non-stressed countries Mean Median Standard deviation Domestic sovereign exposures (%) Non-domestic sovereign exposures (%) Bank lending to firms (%) Bank lending growth (%) Sovereign holdings growth (%) Panel B. Bank characteristics Stressed countries Non-stressed countries Mean Median Standard deviation Mean Median Standard deviation Assets (billion euro) Leverage ratio T1/RWA (%) Deposit/liabilities (%) Public Bailout value as their mean; in the non-stressed countries, it is 6.6%, with a mean of 3.8%. The growth rate of domestic sovereign holdings is more volatile, its standard deviation being 23.1% in stressed countries and 20.1% in non-stressed ones. Both values are very large compared with the respective means of 1.9% and 1%. Both between-banks and withinbank variation in these variables are central to our empirical strategy. Panel B shows that the average bank in the two groups of countries has similar characteristics: it is quite large, highly leveraged (more so in the non-stressed countries), yet with

9 Bank and Sovereign Stress Transmission 2111 Head Banks Domestic Subsidieries Foreign Bank Susidieries Stressed Countries m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 2008m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 2008m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 Head Banks Domestic Subsidieries Foreign Bank Susidieries Non-Stressed Countries m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 2008m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 2008m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 Figure 1. Median domestic sovereign exposure of head banks, domestic and foreign subsidiaries, monthly values. Domestic sovereign exposure is the ratio of domestic sovereign debt holdings to main assets (total assets less derivatives). high regulatory capital ratios (9.4% in the stressed and 9.9% in the non-stressed countries), and mainly reliant on deposit funding (about two-third in both sets of countries). Also, government intervention in the banks of the two groups is similar, with average public stakes of 24% and 23%, respectively (public ownership being defined as shareholdings of local or national government and of publicly controlled institutions); and the frequency of observations referring to bailed-out banks is 10% for both sets of countries (bailout being a dummy equal to 1 during and after a bailout, and 0 otherwise). Figures 1 3 add a dramatic time dimension to two stylized facts that emerge from Table III, namely the rapid growth of banks domestic sovereign exposures and the sharp decline in the loan-to-asset ratio in stressed countries, in striking contrast with the experience of non-stressed countries. Figure 1 shows that the different pattern of sovereign exposures between the two groups of countries is driven by the exposures of the head banks: the median domestic subsidiary in the stressed countries and the median foreign subsidiary in both groups have virtually no sovereign exposures, reflecting the fact that a banking group s securities portfolio is typically managed by the head bank. 5 Figure 2 shows the pattern of median domestic sovereign exposures and loan asset ratios for stressed countries from July 2007 to February 2015; Figure 3 shows the corresponding pattern for non-stressed countries. Besides confirming that domestic sovereign exposures increased much more sharply in stressed countries, the figures illustrate the 5 We are grateful to Rony Hamaui for pointing out this fact to us, based on his managerial experience at Intesa Sanpaolo.

10 2112 C. Altavilla et al. Sovereign exposures (%) Loans/Assets (%) 2008m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 Sovereign exposures (median) Loans/Assets (median) Figure 2. Domestic sovereign exposure and loan asset ratio of the median bank in stressed countries, monthly values. Sovereign exposure is the ratio of domestic sovereign holdings to main assets; loan asset ratio is lending to non-financial corporations divided by main assets. Sovereign exposures (%) Loans/Assets (%) 2008m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 Sovereign exposures (median) Loans/Assets (median) Figure 3. Domestic sovereign exposure and loan asset ratio of the median bank in non-stressed countries, monthly values. Sovereign exposure is the ratio of domestic sovereign holdings to main assets; loan asset ratio is lending to non-financial corporations divided by main assets. completely different dynamics of the median bank s loan-to-asset ratio. Figure 2 shows that in stressed countries, loans to non-financial corporations are correlated negatively with sovereign exposures: over the sample period, the median bank s domestic exposure increases from 1% to 6% of assets, while its corporate lending falls from 28% to about 20% of main assets, the sharpest drop coming in the second half of In late 2014, the loan asset ratio begins to stabilize, in line with the improvement in aggregate lending in the stressed countries. Figure 3 shows a completely different picture for the non-stressed countries: except for the first 2 years of the sample, the loan asset ratio of the median bank is positively

11 Bank and Sovereign Stress Transmission 2113 correlated with its domestic sovereign exposures, and both variables have a distinct positive trend. Of course, these different correlations between sovereign exposures and bank lending at the time-series, aggregate level do not, as such, establish causation: in principle, the negative correlation in stressed countries could reflect either the crowding out of private lending by sovereign debt in banks balance sheets or diminished demand for loans leading banks to substitute them with sovereign debt. However, as we shall see, bank-level data help to pin down the direction of causality, exploiting bank-level heterogeneity in the response of sovereign exposures (Section 3) and of lending (Section 4) to sovereign stress. 3. Determinants of Banks Sovereign Exposures The descriptive evidence set out above highlights the cross-sectional and time-series variation in banks domestic sovereign exposures. Some of this variation is accounted for by three characteristics of the banks: fraction of public share ownership, government bailout history, and regulatory capital ratio. This section documents that these three characteristics correlate not only with differences in sovereign exposures, but also with the way banks vary their exposures when faced with domestic sovereign stress: public ownership, previous occurrence of a bailout, and low capitalization are associated with a greater tendency to increase holdings of distressed government debt in the face of a drop in its price. As observed in Section 1, according to the moral suasion hypothesis, publicly owned banks should be more willing than private ones to surrender to government influence and purchase domestic debt at times of sovereign stress, and foreign banks should be less willing than domestic ones to do so. By the same token, recently rescued banks should be more sensitive to government pressure, their management being typically government-appointed and keenly aware that their survival hinged on a public capital infusion. According to the carry trade hypothesis, poorly capitalized banks should purchase more high-yield public debt to gamble for resurrection. In the case of stressed-country banks, domestic debt is invariably also high-yield debt, so that to distinguish between the two hypotheses heterogeneity across banks is essential: indeed, we exploit the fact that at times of stress public and recently bailed-out banks should be more inclined to buy domestic public debt, and undercapitalized banks to buy more of it for yield-seeking motives. 6 In this section, we show that each of these hypotheses accounts for some of the variation of bank sovereign exposures in stressed countries, and that the two groups of banks to which each hypothesis applies are distinct and largely non-overlapping. Before turning to regression analysis, we provide some graphic evidence to illustrate how changes in domestic sovereign exposures correlate with bank characteristics. Figure 4 shows banks domestic sovereign exposures according to the type of ownership: the lines labeled public and private, respectively, plot the average exposures of banks above and below the average fraction of public share ownership in the relevant country in The two vertical dashed lines in both panels of Figure 4 mark the timing of the two 6 In non-stressed countries, domestic debt obviously does not coincide with high-yield public debt, so that for the banks of those countries one could test the carry trade hypothesis simply by investigating whether they increase their holdings of foreign debt issued by stressed sovereigns. However, our data do not provide a breakdown of foreign sovereign debt holdings by issuer, and therefore prevent us from implementing this test for non-stressed-country banks.

12 2114 C. Altavilla et al. Stressed Countries Non-Stressed Countries m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 2015m1 public private public private Figure 4. Domestic sovereign exposure and bank ownership, in stressed and non-stressed countries. The line labeled public (private) plots the average monthly exposure of banks with a fraction of public ownership above (below) the relevant country average in largest injections of liquidity by the ECB during the sovereign crisis, namely, the 3-year very long-term refinancing operations (VLTROs) of December 2011 and March 2012, which provided loans for e489 bn and e529 bn, respectively, to euro-area banks. 7 In the left panel, which refers to the stressed countries, the domestic sovereign exposures of the two groups of banks are very similar until late 2011, but afterwards the banks with greater public ownership increase their domestic sovereign exposures at a much faster pace than the other group: the difference between them grows from nil in 2011 to over 6 percentage points in 2015, consistently with the moral suasion hypothesis. The largest increase in public banks sovereign exposures occurs in coincidence with the two VLTROs, suggesting that these banks used the liquidity provided by the ECB to fund their purchases of domestic public debt and/or bought such debt to pledge it as collateral to obtain liquidity, as found by Crosignani, Faria-e-Castro, and Fonseca (2016) for Portuguese banks. The right panel shows a qualitatively similar pattern in the domestic exposures of non-stressed countries banks as well, but with a much smaller difference between public and private banks between 1 and 2 percentage points. Figure 5 shows that in stressed countries, banks rescued with public funds purchased substantially more domestic government debt in the month before and the year after it, again consistently with the moral suasion hypothesis. The line plotted in the two panels is the difference between the average domestic sovereign exposure of the bailed-out and the other banks, measured in the same month and group of countries, over a 2-year window centered on the bailout date (month 0). In stressed countries, the exposure of the bailed-out 7 More precisely, the settlement dates of the two operations were December 22, 2011, and March 1, 2012, respectively.

13 Bank and Sovereign Stress Transmission 2115 Stressed Countries Non-Stressed Countries Domestic Sovereign Exposure (%): Bailed-out Minus Control Banks Months to Bail-out Months to Bail-out Figure 5. Difference between the average domestic sovereign exposure of bailed-out and control banks, in stressed and non-stressed countries. Control banks are not bailed-out ones. The difference refers to values observed in the same month and the same group of countries. Month 0 is the bailout date. banks rises on average 3 percentage points above that of the control group over the 12 subsequent months. No such pattern is detectable in non-stressed countries. Figure 6 explores whether banks with lower regulatory capital (Tier-1 capital scaled by risk-weighted assets, or T1/RWA) increased high-yield sovereign holdings more than other banks, consistently with the carry trade hypothesis. The left panel refers to stressed countries, the right panel to non-stressed ones. The figure is based on the subsample of banks for which T1/RWA data are available in the SNL Financial database (SNL): between 30 and 40 banks in each group, depending on month. In each panel, the lines labeled high T1/ RWA and low T1/RWA refer to the average domestic sovereign exposure of banks with above-median and below-median T1/RWA, respectively. After the 2010 Greek bailout, the stressed-country banks with low capital ratios increased their sovereign exposures more than their better-capitalized peers. Some difference, albeit smaller, is also observable in non-stressed countries. Taken together, the three figures suggest that stressed-country banks with more public ownership and less regulatory capital increased their sovereign holdings more than other banks at times of sovereign stress, and recently bailed-out banks bought more stressed domestic debt than other banks. That is, this graphic evidence already suggests that both the moral suasion and the carry trade hypotheses have some explanatory power. Interestingly, the two hypotheses seem to apply to two quite different groups of stressedcountry banks: as of the end of 2008, only one of the low T1/RWA banks in Figure 6 Monte dei Paschi di Siena also features public ownership above its country median, and therefore belongs to the group of public banks in Figure 4.

14 2116 C. Altavilla et al. Stressed Countries Non-Stressed Countries m1 2010m1 2011m1 2012m1 2013m1 2014m1 2009m1 2010m1 2011m1 2012m1 2013m1 2014m1 Low T1/RWA High T1/RWA Low T1/RWA High T1/RWA Figure 6. Domestic sovereign exposure and bank regulatory capital in stressed and non-stressed countries, monthly values. The line labeled High (Low) T1/RWA refers to the average exposure of banks with above-median (below-median) ratio of Tier 1 capital to risk-weighted assets. To test these two hypotheses with regression analysis, we proceed in two steps. Since the SNL data on T1/RWA needed to test the carry trade hypothesis are only available for a subsample of banks, we first use the full sample to test the moral suasion hypothesis only. Next, we restrict the estimation to the subsample for which we have SNL data and test both hypotheses on this smaller sample. In Table IV, we estimate the following specification: DH ijt H ijt 1 ¼ a jt þ c i þ / 1 Public ijt DP jt P jt 1 þ / 2 Public ijt VLTRO t þ / 3 Public ijt þ/ 4 Bailout ijt VLTRO t þ / 5 Bailout ijt þ / 6 F ij DP jt P jt 1 þ/ 7 F ij VLTRO t þ hx ijt 1 þ g ijt ; (1) where the dependent variable is the quarterly percentage change in domestic sovereign holdings H ijt of bank i in country j and quarter t (Holdings H ijt of debt issued by country j s government differ from exposure, which is defined as the ratio of holdings to main assets, i.e., H ijt /A ijt ). In Equation (1), Public it is the time-varying fraction of the bank s shares owned directly or indirectly by local or national government or publicly controlled institutions (Fondazioni in Italy, Fundaciones and Cajas in Spain, and Sparkasse and Landesbank in Germany); DP jt =P jt 1 is the percentage change in the price of sovereign j s debt in the previous quarter (computed as the product of the change in the relevant 10-year yield from t 1 to t by the corresponding duration as in De Marco, 2017); VLTRO t equals 1 in coincidence with the two ECB liquidity injections of December 2011 and March 2012, and 0 otherwise; Bailout ijt equals 1 from the quarter in which bank i was bailed out (unless acquired by

15 Bank and Sovereign Stress Transmission 2117 another bank in the two subsequent quarters), and 0 otherwise; F ij equals 1 if bank i is the subsidiary of a foreign bank operating in country j and 0 if it is a domestic head bank or subsidiary. The specification also includes bank-fixed effects c i to control for unobserved heterogeneity at bank level and time-country effects a jt to control for country-level factors that may affect bank purchases of sovereign debt, including government debt repricing: the latter enters the specification only via its differential effect on banks with different characteristics. Finally, we include the (lagged) deposit liability ratio X ijt 1 as a further bank-level control. In estimating specification (1), errors are clustered at the bank level, and the quarterly growth rates of sovereign holdings are trimmed at 6100% to eliminate outliers. At times of sovereign stress, the price of public debt falls; that is, the variable DP jt =P jt 1 is negative. The moral suasion hypothesis holds that at those times public banks should buy more domestic debt than private ones, and foreign subsidiaries less than domestic banks, so that / 1 < 0 and / 6 > 0. Insofar as the ECB liquidity injections enabled public banks to buy more domestic public debt than private and foreign ones, one would also expect / 2 > 0 and / 7 < 0. The moral suasion hypothesis does not necessarily imply a positive direct effect of public ownership, / 3 : public banks are supposed to be more pliant at times of sovereign stress, not to increase their public debt holdings more than other banks at all times. Instead, the moral suasion hypothesis requires bailed-out banks to buy more sovereign debt during and after their rescue, compared with other banks in the same country and quarter: / 5 > 0. Moreover, if ECB liquidity injections contributed to domestic public debt purchases by bailed-out banks, one should find / 4 > 0. Specification (1) merges elements from the models of moral suasion estimated by De Marco and Macchiavelli (2014); Acharya et al. (2015); Horvath, Huizinga, and Ioannidou (2015); and Ongena, Popov, and van Horen (2016): the first three studies estimate regressions of sovereign exposures on indicators of political control and government support using EBA stress test data; the fourth focuses on measures of foreign ownership using IBSI data for stressed countries. 8 The estimates in Table IV show that for stressed countries the coefficient of the interaction between public ownership and the change of sovereign debt prices (/ 1 ) is negative and significant, and the coefficients of the bailout variable (/ 5 ) and of the interaction between foreign ownership and sovereign price changes (/ 6 ) are both positive, although the latter is imprecisely estimated: all these estimates conform to the predictions of the moral suasion hypothesis. The estimate of / 1 in Column 3 implies that, in response to a 1% decrease in domestic sovereign debt prices, a 100% publicly owned bank (Public ijt ¼ 1) increased its domestic sovereign holdings by 0.35% more than a 100% private bank (Public ijt ¼ 0); the estimate of / 5 instead implies that bailed-out banks increase their public debt holdings by 6.44% more than other banks. Moreover, the interaction of the VLTRO t dummy with public ownership has a positive and significant coefficient (/ 2 ), and that with foreign ownership has a negative and significant one (/ 7 ): the 3-year ECB loans in allowed domestic public banks of stressed countries to purchase sovereign debt far in excess of private and foreign banks. The estimates in Column 2 imply that in the two months of the liquidity injections a 100% publicly owned bank increased its domestic debt holdings 8 The specification used by Ongena, Popov, and van Horen (2016) also relies on a different variable to gauge sovereign stress, namely a measure of abnormally large domestic sovereign issuance (high needs), which may induce the government to pressure domestic banks to underwrite larger amounts of its debt.

16 2118 C. Altavilla et al. Table IV. Determinants of sovereign holdings: moral suasion The dependent variable is the growth rate of banks domestic sovereign holdings in quarter t (defined as the percentage difference between the end-of-period values in quarter t and quarter t 1). The stressed countries are Greece, Ireland, Italy, Portugal, Slovenia, and Spain. The nonstressed countries are Austria, Belgium, Finland, France, Germany, Malta, and the Netherlands. DP jt =P jt 1 is sovereign debt repricing, defined as the percentage change of debt prices in country j and quarter t, based on 10-year benchmark yields. Public ijt is the fraction of banks shares owned by local or national government or publicly controlled institutions (Fondazioni in Italy, Fundaciones and Cajas in Spain, and Sparkasse and Landesbank in Germany). VLTRO equals 1 in December 2011 and March 2012, and 0 otherwise. Bailout ijt equals 1 starting in the quarter t in which bank i in country j was bailed out (unless acquired in the two subsequent quarters), and 0 before quarter t. F ij equals 1 if bank i in country j is a foreign subsidiary and 0 otherwise. All the regressions include the bank-level (lagged) deposit liability ratio as a further control. The sample ranges from 2008:Q1 to 2014:Q4. Standard errors are clustered at the bank level and shown in parentheses: p < 0:01; p < 0:05; p < 0:1. Stressed countries Non-stressed countries (1) (2) (3) (4) (5) (6) Public it DPjt Pjt ** 0.29** 0.35** (0.14) (0.14) (0.15) (0.04) (0.05) (0.05) Public it VLTRO 21.03*** 16.52*** 17.54*** (6.04) (5.92) (5.72) (3.68) (3.95) (4.18) Public it (5.25) (5.13) (6.37) (4.21) (4.14) (6.86) Bailout it VLTRO (5.11) (8.30) Bailout it 6.44** 8.02 (2.65) (6.03) F ij DPjt Pjt * 0.06 (0.11) (0.05) F ij VLTRO 11.98*** 6.83* (4.29) (3.83) Bank FE Yes Yes Yes Yes Yes Yes Time country FE Yes Yes Yes Yes Yes Yes Only domestic No No Yes No No Yes Adjusted R Banks Observations by 16.52% more than those of a 100% privately owned bank, in stressed countries. In contrast, none of the coefficients is significantly different from zero in the non-stressed countries, except for / 7, which is also negative and marginally significant. Since sovereign solvency was seriously questioned by investors only for stressed countries, the results support the moral suasion hypothesis. They broadly agree with the results of De Marco and Macchiavelli (2014) and Horvath, Huizinga, and Ioannidou (2015), but not with those of Acharya et al. (2015), who find no evidence of moral suasion, nor with Ongena, Popov,

17 Bank and Sovereign Stress Transmission 2119 and van Horen (2016), who find no significant interaction between the VLTRO and moral suasion. In Table V, we expand specification (1) to jointly test the moral suasion and the carry trade hypothesis, allowing for their respective interactions with the ECB liquidity injections: DH ijt T1 ¼ a jt þ c H i þ d 1 DP jt T1 T1 þ d 2 VLTRO t þ d 3 þ ijt 1 RWA ijt 1 P jt 1 RWA ijt 1 RWA ijt 1 þd 4 Public ijt DP jt P jt 1 þ d 5 Public ijt VLTRO t þ d 6 Public ijt þd 7 Bailout ijt VLTRO t þ d 8 Bailout ijt : (2) According to the carry trade hypothesis, weakly capitalized banks (low T1=RWA ijt 1 ) should increase their sovereign holdings more than better capitalized ones when government debt becomes cheaper (DP jt =P jt 1 < 0), and resell it more aggressively if and when prices recover (DP jt =P jt 1 > 0) to realize their profits. Hence, the coefficient of the interaction between T1=RWA ijt 1 and DP jt =P jt 1 should be positive: d 1 > 0. Interestingly, the T1=RWA ijt 1 variable has low correlation with Public ijt and Bailout ijt (0.15 and 0.18, respectively), confirming that the group of poorly capitalized banks is quite distinct from the groups of public and recently bailed-out banks. Specification (2) also allows us to test whether weakly capitalized banks borrowed more from the ECB and used these loans to buy risky sovereign debt, as found by Drechsler et al. (2016): this would require the coefficient of the interaction between bank capitalization (T1=RWA ijt 1 )andthevltro t to be negative, that is, d 2 < 0. It is worth noticing that the carry trade hypothesis does not imply that poorly capitalized banks invariably purchase more domestic public debt (i.e., d 3 < 0): if the price of domestic sovereign debt is stable while that of distressed foreign sovereign debt declines, a yield-seeking bank will bet on foreign sovereign debt, and divest domestic debt. In other words, the hypothesis predicts an increasing home bias in sovereign debt portfolios only for banks in stressed countries, not in non-stressed ones: during the crisis, a yield-seeking German bank would not have invested in German but in Italian or Spanish public debt. However, our data only provide a breakdown between domestic and foreign euro-area sovereign debt holdings, and therefore they allow us to test the carry trade hypothesis only for stressed countries: for the banks in non-stressed countries, such testing would require the complete breakdown of their foreign debt portfolio (as in the studies of Buch, Koetter, and Ohls (2016), on German banks and Peydro, Polo, and Sette (2017), on Italian banks). Hence, we estimate Specification (2) only for stressed countries, where our data allow meaningful estimation of the carry-trade coefficients d 1, d 2, and d 3. Specification (2) also includes the variables present in Specification (1) to capture moral suasion, except for the interaction between foreign ownership and sovereign debt repricing, since we have no data on the regulatory capital of foreign subsidiaries. The sample includes only the bank-quarter observations for which the SNL database supplies regulatory capital data. The panel is unbalanced, since there are data gaps even for some of the forty-one banks included in the sample. The estimates of Specification (2) are shown in Table V. The first two columns are for the carry-trade variables only: the sample used in Column 1 includes all domestic banks, while that in Column 2 includes head banks only (that hold most of their groups sovereign debt). The estimate of d 1 is positive and significant in both columns. Its estimate in Column